[Objective] To address the unclear influence mechanisms of multi-point contact states and irregular particle shapes on particle breakage modes and crushing strength in rockfill dams, and to fill the research gap in systematically integrating meso-scale mechanical features with data-driven machine learning algorithms, this study establishes a robust framework for failure mode classification and crushing strength prediction of single rockfill particles. The framework is intended to overcome the inherent limitations of traditional empirical formulas that rely on idealized spherical geometries and fixed two-point elastic contact modes. [Methods] Uniaxial compression tests were conducted on artificial gypsum particles with varying particle sizes, aspect ratios, and loading angles. Digital image correlation (DIC) technique was employed to acquire full-field surface displacement data and capture crack evolution characteristics during compression. A grid search strategy combined with stratified 5-fold cross-validation was applied to optimize hyperparameters and mitigate overfitting. Support Vector Machine (SVM), Multilayer Perceptron (MLP), Random Forest (RF), and eXtreme Gradient Boosting (XGB)—were systematically evaluated and compared for both failure mode classification and crushing strength regression tasks. [Results] Three distinct failure modes were identified: splitting, crushing, and shear. Shear failure occurred exclusively under three-point or higher contact conditions, while two-point compression produced only splitting or crushing. Crushing strength exhibited a pronounced size effect and shape dependency: strength decreased as particle size increased, aspect ratio increased, or loading angle increased. Among the four algorithms, XGB demonstrated the best overall performance. In the classification task, XGB achieved a macro-average AUC of 0.96 on the independent test set, with a macro-average precision of 0.90 and an F1-score of 0.88. In the regression task, XGB yielded the lowest RMSE and MAE of 0.9 MPa and 0.8 MPa, respectively, with R² of 0.92. Compared with four traditional strength criteria (splitting, crushing, shear failure theories, and Hertz contact theory), the XGB model demonstrated superior prediction accuracy across all particle sizes. Under the 20 mm particle size condition, the absolute deviation between the XGB-predicted value and the experimental value was only 0.06 MPa, compared with 0.09 MPa for the Hertz contact theory. [Conclusions] Particle shape and contact state exert significant nonlinear regulatory effects on the crushing characteristics of granular materials. The XGB-based machine learning model effectively overcomes the oversimplified assumptions of traditional theories regarding idealized spherical geometries and fixed contact modes. By autonomously learning from high-dimensional meso-mechanical features, the model exhibits superior generalization ability and prediction accuracy under complex conditions involving irregular particle shapes and multi-point contacts.